hide boss reasoning to avoid injection

support/build_graph.py

@@ -231,7 +231,7 @@ class MyGraph:
         #     entry.get("content", "")
         #     for entry in chat_history
         # ])
-        formatted_history, current_round_messages = self._split_chat_history(chat_history)
+        formatted_history, current_round_messages = self._split_chat_history(chat_history, viewer_role="boss")
         self.current_team = state["current_team"]
 
         new_message = [
@@ -488,7 +488,7 @@ class MyGraph:
         #     entry.get("content", "")
         #     for entry in chat_history
         # ])
-        formatted_history, current_round_messages = self._split_chat_history(chat_history)
+        formatted_history, current_round_messages = self._split_chat_history(chat_history, viewer_role="boss")
 
         boss = next((p for p in state["players"] if p.team == "blue" and p.role == "boss"), None)
         if not boss:
@@ -1212,7 +1212,30 @@ class MyGraph:
         elif hasattr(message, 'text') and message.text:
             content = message.text
         else:
-            content = str(message)
+            # if player.role == "boss":
+            #     content = str(message.content)  # this is to avoid the injection of any reasoning of the boss inside the chat history, that can add information about the words colors on the board
+            # else:
+            model_provider = message.response_metadata.get('model_provider')
+            if (model_provider in ["openai", "google_genai", "anthropic"] and isinstance(message.content, list)):
+                for item in message.content:
+                    if not isinstance(item, dict):
+                        continue
+
+                    item_type = item.get('type')
+                    # Handle openai reasoning content
+                    if item_type == 'reasoning':
+                        content = item.get('summary', [])
+
+                    # Handle Google and Anthropic thinking content
+                    elif item_type == 'thinking':
+                        # reasoning_text = ""
+                        content = item.get('thinking', [])
+
+                    # Handle text content (regular response after reasoning)
+                    elif item_type == 'text':
+                        content = item.get('text', '')
+            elif message.additional_kwargs.get('reasoning_content'):
+                content = message.additional_kwargs.get('reasoning_content')
 
         # Return structured entry
         if player is None:
@@ -1236,14 +1259,15 @@ class MyGraph:
             "info": info,
         }, clue, clue_number, guesses
 
-    def _split_chat_history(self, chat_history):
+    def _split_chat_history(self, chat_history, viewer_role=None):
         """
         Splits chat_history into two parts:
         - formatted_history: all entries up to and including the last one with info == "chat_history"
         - current_round: all entries after that
-
-        Both are formatted strings, where each line is:
-            [name (team role)]: content
+        
+        If viewer_role is specified, filters content appropriately:
+        - 'captain': hides boss reasoning from current round
+        - 'boss' or None: shows everything
         """
 
         last_index = None
@@ -1254,24 +1278,34 @@ class MyGraph:
                 last_index = i
 
         # Helper to format a single entry
-        def format_entry(entry):
+        def format_entry(entry, hide_boss_reasoning=False):
             name = entry.get("name", "Unknown")
             team = entry.get("team", "N/A")
             role = entry.get("role", "N/A")
+
+            # Hide boss reasoning only for captains viewing current round
+            if hide_boss_reasoning and entry.get("role") == "boss" and entry.get("tool_name") is None:
+                return f"[{name} ({team} {role})]: [boss reasoning is hidden]"  # avoid injection of boss reasoning that may contains the association between words and colors
+
             content = entry.get("content", "")
             return f"[{name} ({team} {role})]: {content}"
 
         # Build formatted strings
         if last_index is None:
             formatted_history = ""
-            current_round = "\n".join(format_entry(e) for e in chat_history)
+            current_round_entries = chat_history
         else:
+            # History is always shown as-is (no filtering)
             formatted_history = "\n".join(
-                format_entry(e) for e in chat_history[:last_index + 1]
-            )
-            current_round = "\n".join(
-                format_entry(e) for e in chat_history[last_index + 1:]
+                format_entry(e, hide_boss_reasoning=False) for e in chat_history[:last_index + 1]
             )
+            current_round_entries = chat_history[last_index + 1:]
+
+        # Current round: hide boss reasoning only if viewer is captain
+        hide_reasoning = (viewer_role != "boss")
+        current_round = "\n".join(
+            format_entry(e, hide_boss_reasoning=hide_reasoning) for e in current_round_entries
+        )
 
         return formatted_history, current_round